Fix the optimisation

[jamesrobertlloyd]

This means both

• ensuring local extrema are reached so Laplace approx valid
• jittering previously learned parameter values to avoid having to multiply by constant kernels

[jamesrobertlloyd]

Have experimented with latest version of GPML optimiser - same local extremum problems around - asking Carl. Might be instructive to try the failure example in GPy to see if it exists there as well.